From 468566a1353d8b14f3dd9d6601a0d53d4af7bf71 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sat, 26 Sep 2026 15:22:42 -0500
Subject: [PATCH 1/7] test: rerun hybrid cache-source validation for vLLM
 PR56318

---
 benchmarks/benchmark_lib.sh                   |  22 +-
 .../cache-sources-dep4-dep16-c256-mtp.yaml    | 230 ++++++++++++++++++
 .../dsv4_fp4_b200_vllm_cache_sources_mtp.sh   | 108 ++++++++
 configs/nvidia-master.yaml                    |  53 ++++
 docs/configuration-procedures.md              |   6 +
 docs/configuration-procedures_zh.md           |   6 +
 infx/matrix/generate.py                       |  11 +-
 infx/matrix/validation.py                     |  18 +-
 .../matrix/test_generate_sweep_configs.py     |  28 +++
 infx/tests/matrix/test_validation.py          |  15 ++
 perf-changelog.yaml                           |  12 +
 runners/launch_b200-nscale-slurm.sh           |  36 ++-
 runners/launch_gb300-nv.sh                    |   5 +
 runners/srt-slurm/validation/pr56318.patch    |  24 ++
 utils/test_agentic_offload_modes.py           |  44 ++++
 15 files changed, 598 insertions(+), 20 deletions(-)
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml
 create mode 100644 benchmarks/single_node/agentic/dsv4_fp4_b200_vllm_cache_sources_mtp.sh
 create mode 100644 runners/srt-slurm/validation/pr56318.patch
 create mode 100644 utils/test_agentic_offload_modes.py

diff --git a/benchmarks/benchmark_lib.sh b/benchmarks/benchmark_lib.sh
index b29ad3686e..fd64b96f68 100644
--- a/benchmarks/benchmark_lib.sh
+++ b/benchmarks/benchmark_lib.sh
@@ -192,6 +192,8 @@ require_agentic_kv_offload_none() {
 
 require_agentic_kv_offload_backend() {
     local expected_backend="$1"
+    # Existing recipes support DRAM; callers opt in explicitly to additional tiers.
+    local supported_modes="dram ${2:-}"
     if [[ -z "${KV_OFFLOADING+x}" || -z "$KV_OFFLOADING" ]]; then
         echo "Error: KV_OFFLOADING must be set for agentic benchmarks" >&2
         exit 1
@@ -204,19 +206,23 @@ require_agentic_kv_offload_backend() {
             fi
             return 1
             ;;
-        dram)
+        dram|nvme|dram+nvme)
+            if [[ " $supported_modes " != *" $KV_OFFLOADING "* ]]; then
+                echo "Error: this recipe does not support $KV_OFFLOADING with $expected_backend" >&2
+                exit 1
+            fi
             if [[ "${KV_OFFLOAD_BACKEND:-}" != "$expected_backend" ]]; then
-                echo "Error: expected KV_OFFLOAD_BACKEND=$expected_backend when KV_OFFLOADING=dram, got '${KV_OFFLOAD_BACKEND:-}'" >&2
+                echo "Error: expected KV_OFFLOAD_BACKEND=$expected_backend when KV_OFFLOADING=$KV_OFFLOADING, got '${KV_OFFLOAD_BACKEND:-}'" >&2
                 exit 1
             fi
-            if [[ ! "${TOTAL_CPU_DRAM_GB:-}" =~ ^[1-9][0-9]*$ ]]; then
+            if [[ "$KV_OFFLOADING" != nvme && ! "${TOTAL_CPU_DRAM_GB:-}" =~ ^[1-9][0-9]*$ ]]; then
                 echo "Error: DRAM KV offloading requires a positive TOTAL_CPU_DRAM_GB capacity" >&2
                 exit 1
             fi
             return 0
             ;;
         *)
-            echo "Error: unsupported KV_OFFLOADING value '$KV_OFFLOADING' (expected one of: none, dram)" >&2
+            echo "Error: unsupported KV_OFFLOADING value '$KV_OFFLOADING' (expected none, dram, nvme, or dram+nvme)" >&2
             exit 1
             ;;
     esac
@@ -398,18 +404,18 @@ if [[ "$_benchmark_caller" == */agentic/* ||
                 exit 1
             fi
             ;;
-        dram)
+        dram|nvme|dram+nvme)
             if [[ -z "${KV_OFFLOAD_BACKEND:-}" || "${KV_OFFLOAD_BACKEND:-}" == "none" ]]; then
-                echo "Error: KV_OFFLOAD_BACKEND is required when KV_OFFLOADING=dram" >&2
+                echo "Error: KV_OFFLOAD_BACKEND is required when KV_OFFLOADING=$KV_OFFLOADING" >&2
                 exit 1
             fi
-            if [[ ! "${TOTAL_CPU_DRAM_GB:-}" =~ ^[1-9][0-9]*$ ]]; then
+            if [[ "$KV_OFFLOADING" != nvme && ! "${TOTAL_CPU_DRAM_GB:-}" =~ ^[1-9][0-9]*$ ]]; then
                 echo "Error: DRAM KV offloading requires a positive configured TOTAL_CPU_DRAM_GB capacity" >&2
                 exit 1
             fi
             ;;
         *)
-            echo "Error: unsupported KV_OFFLOADING value '$KV_OFFLOADING' (expected one of: none, dram)" >&2
+            echo "Error: unsupported KV_OFFLOADING value '$KV_OFFLOADING' (expected none, dram, nvme, or dram+nvme)" >&2
             exit 1
             ;;
     esac
diff --git a/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml
new file mode 100644
index 0000000000..853ca3f404
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml
@@ -0,0 +1,230 @@
+# PR56318 cache-source validation: preserve run34406466616 resources.
+# One DEP4 prefill worker/node and one DEP16 decode worker across four nodes.
+schema: 2
+name: "dsv4-vllm-cache-sources-gb300-dep4-dep16-c256"
+
+# GB300 AgentX MTP3 topology: one DEP4 prefill worker feeds one
+# DEP16 decode worker at concurrency 256. Decode consumes P/D KV through NIXL
+# and MooncakeStore but skips Mooncake prefix lookup to avoid CPU overhead.
+
+model:
+  path: "deepseek-v4-pro"
+  container: "cquil11/vllm-cache-sources@sha256:6ce87fcb9ef8b929f7a8270bc6fbba3c219c3440cd77d3bf1ae74db5a1cb5715"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "cquil11/vllm-cache-sources@sha256:6ce87fcb9ef8b929f7a8270bc6fbba3c219c3440cd77d3bf1ae74db5a1cb5715"
+  frameworks:
+    dynamo: "1.4.0"
+
+dynamo:
+  install: true
+
+  source:
+    wheel: "1.4.0"
+environment:
+  # Mooncake prefix-block hashes must match across processes and nodes.
+  PYTHONHASHSEED: "0"
+
+setup_script: vllm-container-deps.sh
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 2160
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  het_jobs: false
+  spread_workers: false
+services:
+  - name: etcd
+    type: etcd
+    placement:
+      node: infra
+  - name: nats
+    type: nats
+    placement:
+      node: infra
+    options:
+      max_payload_mb: 32
+  - name: mooncake-master
+    type: mooncake-master
+    options:
+      store_config:
+        metadata_server: "P2PHANDSHAKE"
+        global_segment_size: "180GB"
+        local_buffer_size: "4GB"
+        protocol: "rdma"
+        device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+        mode: "embedded"
+        enable_offload: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    router-mode: "random"
+    router-session-affinity-ttl-secs: 900
+  env:
+    DYN_ROUTER_ACTIVE_REQUEST_EXPIRY_SECS: "3600"
+    DYN_TCP_CHANNEL_BUFFER: "128"
+    DYN_TCP_REQUEST_TIMEOUT: "60"
+
+engine:
+  type: vllm
+  connector:
+  dp_launch_mode: per_node
+roles:
+  prefill:
+    nodes: 1
+    workers: 1
+    gpus: 4
+    env:
+      HF_HUB_CACHE: "/hf_hub_cache"
+      HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+      TRANSFORMERS_CACHE: "/hf_hub_cache"
+      VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+      VLLM_RPC_TIMEOUT: "600000"
+      VLLM_LOG_STATS_INTERVAL: "1"
+      VLLM_V2_WARMUP_MAX_NUM_SEQS: "20"
+      VLLM_SERVER_DEV_MODE: "1"
+      VLLM_USE_V2_MODEL_RUNNER: "1"
+      VLLM_USE_RUST_FRONTEND: "0"
+      VLLM_MOONCAKE_LOAD_RECV_THREADS: "20"
+      VLLM_MOONCAKE_STORE_SEND_THREADS: "8"
+      VLLM_ALLREDUCE_USE_SYMM_MEM: "0"
+      UCX_MEMTYPE_CACHE: "n"
+      UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+      UCX_TLS: "rc,cuda_copy"
+      NCCL_CUMEM_ENABLE: "1"
+      NCCL_MNNVL_ENABLE: "1"
+      NCCL_NVLS_ENABLE: "1"
+      VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
+      NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "0"
+      VLLM_USE_BREAKABLE_CUDAGRAPH: "1"
+      VLLM_CONNECTOR_PREFETCH_DEPTH: "8"
+      VLLM_DSV4_MEGA_FP8_COMBINE: "1"
+      DG_JIT_CACHE_DIR: "/tmp/dg-cache-dsv4-gb300-1p1d-dep8-prefill-c256-{job_id}"
+      MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
+      MC_STORE_CLIENT_METRIC: "1"
+      MC_STORE_CLIENT_METRIC_INTERVAL: "5"
+      MC_TE_METRIC: "0"
+    args:
+      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      safetensors-load-strategy: "prefetch"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-cumem-allocator: true
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      max-model-len: 1048576
+      max-num-seqs: 256
+      max-num-batched-tokens: 8192
+      long-prefill-token-threshold: 1024
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      gpu-memory-utilization: 0.92
+      no-disable-hybrid-kv-cache-manager: true
+      tokenizer-mode: "deepseek_v4"
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"indexer_kv_dtype":"mxfp4"}'
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
+      moe-backend: "deep_gemm_mega_moe"
+      numa-bind: true
+      numa-bind-nodes: [0, 0, 1, 1]
+  decode:
+    nodes: 4
+    workers: 1
+    gpus: 16
+
+    env:
+      HF_HUB_CACHE: "/hf_hub_cache"
+      HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+      TRANSFORMERS_CACHE: "/hf_hub_cache"
+      VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+      VLLM_RPC_TIMEOUT: "600000"
+      VLLM_LOG_STATS_INTERVAL: "1"
+      VLLM_V2_WARMUP_MAX_NUM_SEQS: "20"
+      VLLM_SERVER_DEV_MODE: "1"
+      VLLM_USE_V2_MODEL_RUNNER: "1"
+      VLLM_USE_RUST_FRONTEND: "0"
+      VLLM_MOONCAKE_LOAD_RECV_THREADS: "4"
+      VLLM_ALLREDUCE_USE_SYMM_MEM: "0"
+      UCX_MEMTYPE_CACHE: "n"
+      UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+      UCX_TLS: "rc,cuda_copy"
+      NCCL_CUMEM_ENABLE: "1"
+      NCCL_MNNVL_ENABLE: "1"
+      NCCL_NVLS_ENABLE: "1"
+      VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
+      NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "0"
+      VLLM_DSV4_MEGA_FP8_COMBINE: "1"
+      DG_JIT_CACHE_DIR: "/tmp/dg-cache-dsv4-gb300-1p1d-dep32-decode-c256-{job_id}"
+      MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
+      MC_STORE_CLIENT_METRIC: "1"
+      MC_STORE_CLIENT_METRIC_INTERVAL: "5"
+      MC_TE_METRIC: "0"
+
+    args:
+      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_consumer","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_consumer","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_consumer","kv_connector_extra_config":{"load_async":true,"lookup_async":false,"enable_lookup":false,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      safetensors-load-strategy: "prefetch"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 16
+      data-parallel-rpc-port: 13345
+      enable-cumem-allocator: true
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      max-model-len: 1048576
+      max-num-seqs: 8
+      max-num-batched-tokens: 32
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      max-cudagraph-capture-size: 32
+      gpu-memory-utilization: 0.90
+      no-disable-hybrid-kv-cache-manager: true
+      tokenizer-mode: "deepseek_v4"
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"indexer_kv_dtype":"mxfp4"}'
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
+      moe-backend: "deep_gemm_mega_moe"
+      numa-bind: true
+      numa-bind-nodes: [0, 0, 1, 1]
+sbatch_directives:
+  cpus-per-task: "72"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:prompt_tokens_cached_by_source"
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    # Avoid concurrent readers observing a mismatched mmap data/index pair.
+    AIPERF_DATASET_MMAP_CACHE_ENABLED: "false"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/single_node/agentic/dsv4_fp4_b200_vllm_cache_sources_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_b200_vllm_cache_sources_mtp.sh
new file mode 100644
index 0000000000..0c6a61e9c3
--- /dev/null
+++ b/benchmarks/single_node/agentic/dsv4_fp4_b200_vllm_cache_sources_mtp.sh
@@ -0,0 +1,108 @@
+#!/usr/bin/env bash
+set -eo pipefail
+set -x
+
+# The original DeepSeek-V4-Pro TP8/MTP3 cache-source validation recipe.
+# Preserve run 34420875287's cache pressure independently of the 0813 DSpark curve.
+source "$(dirname "$0")/../../benchmark_lib.sh"
+check_env_vars MODEL MODEL_PATH TP CONC KV_OFFLOADING KV_OFFLOAD_BACKEND
+check_env_vars TOTAL_CPU_DRAM_GB RESULT_DIR DURATION PORT GPU_MEMORY_UTILIZATION
+check_env_vars EP_SIZE DP_ATTENTION DCP_SIZE PCP_SIZE EVAL_ONLY THINKING_MODE
+
+if [[ "$TP" != 8 || "$EP_SIZE" != 1 || "$DP_ATTENTION" != false || "$DCP_SIZE" != 1 || "$PCP_SIZE" != 1 ]]; then
+    echo "Cache-source validation requires TP8 without EP, DP attention, or context parallelism" >&2
+    exit 1
+fi
+export GPU_COUNT=8
+[[ -d "$MODEL_PATH" ]] || { echo "Missing pre-staged model: $MODEL_PATH" >&2; exit 1; }
+nvidia-smi
+resolve_trace_source
+install_agentic_deps
+
+export AIPERF_SERVER_METRICS_URLS="http://localhost:$PORT/metrics"
+export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="vllm:prompt_tokens_cached_by_source"
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+export VLLM_PREFIX_CACHE_RETENTION_INTERVAL=32768
+export VLLM_USE_V2_MODEL_RUNNER=1
+export VLLM_USE_RUST_FRONTEND=0
+export VLLM_DSV4_MEGA_FP8_COMBINE=1
+export VLLM_RPC_TIMEOUT=600000
+export PYTHONHASHSEED=42
+export TORCH_CUDA_ARCH_LIST=10.0
+export PYTHONNOUSERSITE=1
+export VLLM_FLOAT32_MATMUL_PRECISION=high
+
+mkdir -p "$RESULT_DIR"
+SERVER_LOG="$RESULT_DIR/server.log"
+case "$KV_OFFLOAD_BACKEND" in
+    vllm-native)
+        require_agentic_kv_offload_backend vllm-native "dram dram+nvme"
+        SECONDARY_TIERS='[]'
+        if [[ "$KV_OFFLOADING" == dram+nvme ]]; then
+            check_env_vars NVME_OFFLOAD_DIR
+            SECONDARY_TIERS="[{\"type\":\"fs\",\"root_dir\":\"$NVME_OFFLOAD_DIR\",\"locality\":\"LOCAL\"}]"
+        fi
+        OFFLOAD_CONFIG="{\"kv_connector\":\"OffloadingConnector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"spec_name\":\"TieringOffloadingSpec\",\"cpu_bytes_to_use\":$((TOTAL_CPU_DRAM_GB * 1000000000)),\"secondary_tiers\":$SECONDARY_TIERS}}"
+        ;;
+    vllm-simple)
+        require_agentic_kv_offload_backend vllm-simple nvme
+        check_env_vars NVME_OFFLOAD_DIR
+        OFFLOAD_CONFIG="{\"kv_connector\":\"SimpleCPUOffloadConnector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"kv_offload_backend\":\"disk\",\"disk_path\":\"$NVME_OFFLOAD_DIR/cache.bin\",\"disk_capacity_bytes\":$((1000000000000 / GPU_COUNT)),\"disk_buffer_slots\":4,\"lazy_offload\":false}}"
+        ;;
+    *) echo "Unsupported validation offload backend: $KV_OFFLOAD_BACKEND" >&2; exit 1 ;;
+esac
+
+# Use the same committed golden MTP curve as the shared SRT selector.
+SPEC_CONFIG=$(python3 - <<'PY'
+import json
+import os
+from infx.srt_slurm.synthetic_acceptance import GOLDEN_DIR, golden_length
+
+spec = {"method": "mtp", "num_speculative_tokens": 3}
+if os.environ["EVAL_ONLY"] != "true":
+    spec["rejection_sample_method"] = "synthetic"
+    spec["synthetic_acceptance_length"] = golden_length(
+        "dsv4", spec, os.environ["THINKING_MODE"], GOLDEN_DIR
+    )
+print(json.dumps(spec))
+PY
+)
+MAX_NUM_SEQS=$((2 * CONC))
+CAPTURE_SIZE_LIST=()
+for ((num_seqs = 1; num_seqs <= MAX_NUM_SEQS; num_seqs++)); do
+    CAPTURE_SIZE_LIST+=("$((num_seqs * 4))")
+done
+CAPTURE_SIZE_LIST+=(100 200 300 400 500)
+CUDA_GRAPH_CAPTURE_SIZES=$(printf '%s\n' "${CAPTURE_SIZE_LIST[@]}" | sort -n -u | paste -sd, -)
+COMPILATION_CONFIG="{\"cudagraph_mode\":\"FULL_AND_PIECEWISE\",\"cudagraph_capture_sizes\":[${CUDA_GRAPH_CAPTURE_SIZES}]}"
+
+{ set +x; } 2>/dev/null
+VLLM_CMD=(
+    vllm serve "$MODEL_PATH" --served-model-name "$MODEL"
+    --host 0.0.0.0 --port "$PORT" --trust-remote-code
+    --kv-cache-dtype fp8 --block-size 256 --max-model-len 1048576
+    --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION"
+    --numa-bind --enable-cumem-allocator --no-enable-flashinfer-autotune
+    --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4
+    --enable-auto-tool-choice --reasoning-parser deepseek_v4
+    --attention-config '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"indexer_kv_dtype":"mxfp4"}'
+    --speculative-config "$SPEC_CONFIG" --no-disable-hybrid-kv-cache-manager
+    --disable-uvicorn-access-log --compilation-config "$COMPILATION_CONFIG"
+    --max-num-seqs "$MAX_NUM_SEQS" --tensor-parallel-size "$TP" --data-parallel-size 1
+    --kv-transfer-config "$OFFLOAD_CONFIG"
+)
+printf '%q ' "${VLLM_CMD[@]}" | tee "$RESULT_DIR/vllm_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/vllm_command.txt"
+sha256sum -c /opt/pr56318-validation/validation-overlay.sha256 | tee "$RESULT_DIR/overlay-verification.log"
+"${VLLM_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+if [[ "$EVAL_ONLY" == true ]]; then
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    curl --fail --silent --show-error "$AIPERF_SERVER_METRICS_URLS" > "$RESULT_DIR/metrics-before.prom"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+    curl --fail --silent --show-error "$AIPERF_SERVER_METRICS_URLS" > "$RESULT_DIR/metrics-after.prom"
+fi
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index f442fed05b..30d7264832 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8385,6 +8385,59 @@ dsv41flash-fp4-b200-sglang-agentic-dspark:
       - { tp: 4, ep: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/b200-fp4-mtp/agentic.yaml }
       - { tp: 2, ep: 2, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/b200-fp4-mtp/agentic.yaml }
 
+# PR56318: rerun the four successful hybrid cache-source validation points.
+dsv4-fp4-b200-vllm-agentic-cache-sources-mtp:
+  image: cquil11/vllm-cache-sources@sha256:6f9e476f6a1e1c25ac9bc1a863329538db465812fb4ac18c6553e988eae09a56
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: cluster:b200-nscale
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.0592
+      search-space:
+      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: vllm-native }, spec-decoding: mtp, conc-list: [8] }
+    - search-space:
+      - { tp: 8, kv-offloading: nvme, kv-offload-backend: { name: vllm-simple }, spec-decoding: mtp, conc-list: [14] }
+    - dram-utilization: 0.0592
+      search-space:
+      - { tp: 8, kv-offloading: [dram, nvme], kv-offload-backend: { name: vllm-native }, spec-decoding: mtp, conc-list: [14] }
+
+dsv4-fp4-gb300-dynamo-vllm-agentic-cache-sources-mtp-disagg:
+  image: cquil11/vllm-cache-sources@sha256:6ce87fcb9ef8b929f7a8270bc6fbba3c219c3440cd77d3bf1ae74db5a1cb5715
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: cluster:gb300-nv
+  precision: fp4
+  framework: dynamo-vllm
+  kv-p2p-transfer: nixl
+  multinode: true
+  disagg: true
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.857143
+      search-space:
+      - spec-decoding: mtp
+        conc-list: [256]
+        kv-offloading: dram
+        kv-offload-backend: { name: mooncake, version: "0.3.13.post1" }
+        router: { name: dynamo-router, version: "1.4.0" }
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "SLURM_PARTITION=batch_1"
+          - "CONFIG_FILE=recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+
 dsv41flash-fp4-gb300-vllm-agentic-dspark:
   image: vllm/vllm-openai:nightly-af1c01499b289be555c475669ba50a88e96d846e
   model: deepseek-ai/DeepSeek-V4.1-Flash
diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index f57f08c75b..9eb8d1d0ac 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -23,6 +23,12 @@ Use this page for benchmark configuration, recipe, image, and runner changes. It
 
 Delete retired entries from the active master configs; they are not archived. Git history and `perf-changelog.yaml` keep the historical settings. For partial retirements, remove only the retired scenarios. Delete retired AMD server-registry entries and model-specific setup from `benchmarks/multi_node/amd_utils/` as well. Preserve shared dependencies needed by retained SPEED-Bench collectors, including their scheduling scores. See the [deprecation rules](../AGENTS.md#deprecating-benchmark-configs).
 
+## Cache-source validation
+
+Single-node AgentX configs accept `kv-offloading: nvme` or `[dram, nvme]` in addition to `none` and `dram`. Tiered entries still require `dram-utilization`, which budgets host memory only. The dedicated DeepSeek-V4-Pro B200 validation recipe supports Simple NVMe and native DRAM/NVMe. Its launcher mounts a job-owned `/scratch/inferencex-kv-<job-id>` directory and removes it before releasing the allocation. Other recipes must explicitly opt in to NVMe modes.
+
+The dedicated GB300 cache-source recipe applies a job-local srt-slurm patch to scrape every physical DP worker, including nonleader nodes. It verifies the image's PR-overlay checksums after dependency installation. This patch does not run for other recipes.
+
 ## Dependency submodules
 
 Git records the exact dependency commits. [`.gitmodules`](../.gitmodules) defines the repositories: AIPerf at `utils/aiperf`, NVIDIA srt-slurm at `utils/srt-slurm`. TileRT is a documented manual fork checkout in `setup_srt_slurm()`, not a separate submodule.
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index 968553d8a3..1ae3b4b181 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -23,6 +23,12 @@
 
 退役的配置项直接从启用的主配置中删除，不再归档；历史设置由 Git 历史和 `perf-changelog.yaml` 保留。仅弃用部分场景时，只删除已退役的场景。退役的 AMD 服务注册项和模型专用初始化逻辑也应从 `benchmarks/multi_node/amd_utils/` 中删除。保留 SPEED-Bench 采集器仍需使用的共享依赖，包括调度评分。参见[弃用规则](../AGENTS.md#deprecating-benchmark-configs)。
 
+## 缓存来源验证
+
+单节点 AgentX 配置除了 `none` 和 `dram`，还接受 `kv-offloading: nvme` 或 `[dram, nvme]`。分层配置仍须提供 `dram-utilization`，且该值仅用于分配主机内存。专用 DeepSeek-V4-Pro B200 验证配方支持 Simple NVMe 和原生 DRAM/NVMe。launcher 挂载作业专属的 `/scratch/inferencex-kv-<job-id>` 目录，并在释放资源前将其删除。其他配方必须显式启用 NVMe 模式。
+
+专用 GB300 缓存来源配方在作业本地的 srt-slurm 副本中应用补丁，采集每个物理 DP worker 的指标，包括非 leader 节点。依赖安装后会验证镜像中 PR 覆盖文件的校验和。其他配方不会应用此补丁。
+
 ## 依赖子模块
 
 Git 记录依赖的精确提交版本。[`.gitmodules`](../.gitmodules) 定义各仓库：AIPerf 位于 `utils/aiperf`，NVIDIA srt-slurm 位于 `utils/srt-slurm`。TileRT 由 `setup_srt_slurm()` 手动检出已记录的分支仓库，不是独立子模块。
diff --git a/infx/matrix/generate.py b/infx/matrix/generate.py
index 28d0f2a1f3..0349995731 100644
--- a/infx/matrix/generate.py
+++ b/infx/matrix/generate.py
@@ -441,7 +441,7 @@ def agentic_dram_offload_gb(
       budgeted separately if it ever gains its own pool).
     """
     kv_offloading = benchmark.get(Fields.KV_OFFLOADING.value, "none")
-    if kv_offloading != "dram":
+    if kv_offloading not in ("dram", ["dram", "nvme"]):
         return 0
 
     available_mib = min(
@@ -472,13 +472,14 @@ def agentic_dram_offload_gb(
 
 
 def agentic_kv_offload_suffix(
-    kv_offloading: str,
+    kv_offloading: str | list[str],
     kv_offload_backend: dict | None,
 ) -> str:
     """Return a compact exp-name suffix for agentic KV offload settings."""
     if kv_offloading == "none":
         return "kvnone"
-    return f"kv{kv_offloading}-{kv_offload_backend['name']}"
+    mode = "+".join(kv_offloading) if isinstance(kv_offloading, list) else kv_offloading
+    return f"kv{mode}-{kv_offload_backend['name']}"
 
 
 def multinode_agentic_exp_name(
@@ -998,7 +999,9 @@ def _agentic_entries(
             )
         entry.update(
             {
-                Fields.KV_OFFLOADING.value: kv_offloading,
+                Fields.KV_OFFLOADING.value: (
+                    "+".join(kv_offloading) if isinstance(kv_offloading, list) else kv_offloading
+                ),
                 Fields.TOTAL_CPU_DRAM_GB.value: total_cpu_dram_gb,
                 Fields.DURATION.value: DEFAULT_AGENTIC_DURATION_SECONDS,
                 Fields.EXP_NAME.value: exp_name,
diff --git a/infx/matrix/validation.py b/infx/matrix/validation.py
index bf1af5f272..3d432e516a 100644
--- a/infx/matrix/validation.py
+++ b/infx/matrix/validation.py
@@ -14,6 +14,7 @@
 
 CLUSTER_LABEL_PREFIX = "cluster:"
 DEFAULT_AGENTIC_DURATION_SECONDS = 3600
+type KVOffloadingConfig = Literal["none", "dram", "nvme"] | list[Literal["dram", "nvme"]]
 
 """
     The below class defines the field names expected to be present in the JSON entries
@@ -325,7 +326,9 @@ class SingleNodeAgenticMatrixEntry(BaseModel):
         default="none", alias=Fields.SPEC_DECODING.value
     )
     conc: int
-    kv_offloading: Literal["none", "dram"] = Field(alias=Fields.KV_OFFLOADING.value)
+    kv_offloading: Literal["none", "dram", "nvme", "dram+nvme"] = Field(
+        alias=Fields.KV_OFFLOADING.value
+    )
     kv_offload_backend: KVOffloadBackendMetadata | None = Field(
         default=None, alias=Fields.KV_OFFLOAD_BACKEND.value
     )
@@ -516,7 +519,10 @@ def _validate_kv_offload_fields(self: Any) -> Any:
                 f"{Fields.KV_OFFLOAD_BACKEND.value} requires {Fields.KV_OFFLOADING.value}"
             )
         return self
-    if self.kv_offloading == "none":
+    if isinstance(self.kv_offloading, list):
+        if self.kv_offloading != ["dram", "nvme"]:
+            raise ValueError("The only supported tier list is ['dram', 'nvme']")
+    elif self.kv_offloading == "none":
         if backend is not None:
             raise ValueError(
                 f"{Fields.KV_OFFLOAD_BACKEND.value} can only be set when "
@@ -642,9 +648,7 @@ class AgenticCodingSearchSpaceEntry(BaseModel):
     prefill: WorkerConfig | None = None
     decode: WorkerConfig | None = None
     num_nodes: int | None = Field(default=None, alias=Fields.NUM_NODES.value, gt=0, strict=True)
-    kv_offloading: Literal["none", "dram"] | None = Field(
-        default=None, alias=Fields.KV_OFFLOADING.value
-    )
+    kv_offloading: KVOffloadingConfig | None = Field(default=None, alias=Fields.KV_OFFLOADING.value)
     kv_offload_backend: KVOffloadBackendMetadata | None = Field(
         default=None, alias=Fields.KV_OFFLOAD_BACKEND.value
     )
@@ -690,6 +694,8 @@ def validate_topology_fields(self) -> Self:
                 )
             _validate_tp_context_topology(self)
         if has_aggregate_worker or has_complete_multinode:
+            if self.kv_offloading in ("nvme", ["dram", "nvme"]):
+                raise ValueError("NVMe offloading currently requires a single-node entry")
             explicitly_single_node_fields = {
                 "pp",
                 "dcp_size",
@@ -725,7 +731,7 @@ class AgenticCodingConfig(BaseModel):
     @model_validator(mode="after")
     def validate_dram_offload_capacity(self) -> Self:
         for entry in self.search_space:
-            if entry.kv_offloading != "dram":
+            if entry.kv_offloading not in ("dram", ["dram", "nvme"]):
                 continue
             if self.dram_utilization is None:
                 raise ValueError(
diff --git a/infx/tests/matrix/test_generate_sweep_configs.py b/infx/tests/matrix/test_generate_sweep_configs.py
index e119694d80..f15feb2fe6 100644
--- a/infx/tests/matrix/test_generate_sweep_configs.py
+++ b/infx/tests/matrix/test_generate_sweep_configs.py
@@ -2495,6 +2495,34 @@ def agentic_config(request, sample_single_node_config):
 
 
 class TestAgenticGeneration:
+    @pytest.mark.parametrize(("mode", "runtime", "budget"), [
+        ("dram", "dram", 1199),
+        ("nvme", "nvme", 0),
+        (["dram", "nvme"], "dram+nvme", 1199),
+    ])
+    def test_offload_modes_preserve_budget_and_artifact_identity(
+        self, sample_single_node_config, sample_runner_config,
+        generate_agentic_sweep, mode, runtime, budget,
+    ):
+        config = copy.deepcopy(sample_single_node_config)
+        entry = next(iter(config.values()))
+        entry.update(runner="cluster:b300-nv", multinode=False)
+        entry["scenarios"] = {"agentic-coding": [{
+            "dram-utilization": 0.80,
+            "search-space": [{
+                "tp": 4, "kv-offloading": mode,
+                "kv-offload-backend": {"name": "vllm-native"}, "conc-list": [8],
+            }],
+        }]}
+        original = copy.deepcopy(config)
+        rows = generate_agentic_sweep(config, sample_runner_config)
+        assert rows
+        for row in rows:
+            assert row["kv-offloading"] == runtime
+            assert row["total-cpu-dram-gb"] == budget
+            assert f"kv{runtime}-vllm-native" in row["exp-name"]
+        assert config == original
+
     def test_point_order_and_input_preservation(
         self, agentic_config, sample_runner_config, generate_agentic_sweep,
     ):
diff --git a/infx/tests/matrix/test_validation.py b/infx/tests/matrix/test_validation.py
index 62db6815e0..768296f8d6 100644
--- a/infx/tests/matrix/test_validation.py
+++ b/infx/tests/matrix/test_validation.py
@@ -301,6 +301,21 @@ def test_disagg_requires_multinode(self, valid_single_node_matrix_entry):
 class TestAgenticMatrixEntries:
     """Tests for agentic coding validation models."""
 
+    @pytest.mark.parametrize("mode", [[], ["dram"], ["nvme", "dram"], ["dram", "dram"]])
+    def test_rejects_unsupported_tier_lists(self, mode):
+        with pytest.raises(ValidationError, match="only supported tier list"):
+            AgenticCodingSearchSpaceEntry(**{
+                "tp": 8, "kv-offloading": mode,
+                "kv-offload-backend": {"name": "vllm-native"}, "conc-list": [8],
+            })
+
+    def test_tiered_offload_requires_dram_budget(self):
+        with pytest.raises(ValidationError, match="dram-utilization"):
+            AgenticCodingConfig(**{"search-space": [{
+                "tp": 8, "kv-offloading": ["dram", "nvme"],
+                "kv-offload-backend": {"name": "vllm-native"}, "conc-list": [8],
+            }]})
+
     def test_arbitrary_backend_is_valid_for_single_node_agentic_entry(self):
         entry = SingleNodeAgenticMatrixEntry(**{
             "image": "cquil/vllm-openai:v0.21.0-8813c92",
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0f361fd6db..04bd39c76e 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8962,3 +8962,15 @@
     - "旧镜像切出于 2026-09-16，新镜像领先 441 个提交，包含当前曲线所不含的三项 Kimi-K3 正确性修复：vllm-project/vllm#51483 不再将无状态的首个 chunk 当作 decode，vllm-project/vllm#57098 修复 Kimi-K3 reasoning parser，vllm-project/vllm#57430 支持 routed expert 量化。新镜像还包含本臂在并发大于 4 时均会使用的 ROCm CPU KV 卸载改动，包括 vllm-project/vllm#57160（ROCm CPU KV 卸载改用私有 pinned 张量）与 vllm-project/vllm#50045（卸载背压检测）。"
     - "This image bump does not change the DSpark draft model data type. Only the image: line changes and kimik3_fp4_mi355x_mtp.sh is unchanged; the draft loads unmodified from the published Inferact/Kimi-K3-DSpark checkpoint via --speculative-config (model=Inferact/Kimi-K3-DSpark, method=dspark). The only dtype in that speculative-config is kv_cache_dtype=fp8, which sets the draft KV-cache storage precision, not the draft weights. No flag overrides or re-quantizes the draft-model weights, so the draft dtype is preserved from its checkpoint across this re-sweep."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3419
+
+- config-keys:
+    - dsv4-fp4-b200-vllm-agentic-cache-sources-mtp
+    - dsv4-fp4-gb300-dynamo-vllm-agentic-cache-sources-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  no-evals: true
+  description:
+    - "Rerun vLLM PR56318 at bd57138f9b98eb64c44ea6a0f10d08c8f639820e with the four proven DeepSeek-V4-Pro hybrid cache-source points: B200 TP8 MTP3 native DRAM c8, Simple NVMe c14, and native DRAM+NVMe c14; GB300 MTP3 NIXL plus Mooncake c256."
+    - "Preserve B200 GPU utilization 0.85, native DRAM utilization 0.0592, the dedicated 1 TB Simple NVMe budget, 3600-second AgentX profiles, and disabled evals. GB300 is one DEP4 prefill worker on one node plus one DEP16 decode worker across four nodes: five nodes and twenty GPUs total."
+    - "Keep existing benchmark curves unchanged. Adapt the isolated old-Pro validation recipes to current InferenceX paths and current vLLM flag names, use the Python frontend, and collect all raw backend metrics with the cached-prompt source counter required."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
index 1c6c7afc04..291538f32b 100755
--- a/runners/launch_b200-nscale-slurm.sh
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -786,6 +786,10 @@ run_agentic() {
     if [[ ! -f "$BENCH_SCRIPT" ]]; then
         BENCH_SCRIPT="${BENCH_BASE}${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh"
     fi
+    if [[ "$MODEL" == "deepseek-ai/DeepSeek-V4-Pro" && "$FRAMEWORK" == vllm && "$SPEC_DECODING" == mtp && "$IMAGE" == cquil11/vllm-cache-sources@sha256:6f9e476f6a1e1c25ac9bc1a863329538db465812fb4ac18c6553e988eae09a56 ]]; then
+        BENCH_SCRIPT="benchmarks/single_node/agentic/dsv4_fp4_b200_vllm_cache_sources_mtp.sh"
+        export GPU_MEMORY_UTILIZATION=0.85
+    fi
     LOCK_FILE="${SQUASH_FILE}.lock"
 
     # TODO(Cam): lmsysorg/sglang:deepseek-v4-blackwell installs sglang editable at
@@ -793,7 +797,7 @@ run_agentic() {
     # the default $GITHUB_WORKSPACE:/workspace/ bind-mount masks the install and
     # breaks `import sglang`. Mount this one image at /ix instead; drop the
     # conditional once the image stops installing editable under /workspace.
-    if [[ "$IMAGE" == *deepseek-v4-blackwell* ]]; then
+    if [[ "$IMAGE" == *deepseek-v4-blackwell* || "$BENCH_SCRIPT" == *cache_sources_mtp.sh ]]; then
         CONTAINER_MOUNT_DIR=/ix
     else
         CONTAINER_MOUNT_DIR=/workspace
@@ -810,9 +814,37 @@ run_agentic() {
     else
         CONTAINER_MOUNTS="$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR,$MODEL_PATH:$MODEL_PATH,$AIPERF_MMAP_CACHE_HOST_PATH:/aiperf_mmap_cache"
     fi
+    if [[ "$BENCH_SCRIPT" == *cache_sources_mtp.sh ]]; then
+        export INFMAX_CONTAINER_WORKSPACE="$CONTAINER_MOUNT_DIR"
+        export RESULT_DIR="$CONTAINER_MOUNT_DIR/results"
+    fi
 
-    salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --gres=gpu:$GPU_COUNT --exclusive --mem=0 --time="$SALLOC_TIME_LIMIT" --no-shell --job-name="$RUNNER_NAME"
+    salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --gres=gpu:$GPU_COUNT --exclusive --mem=0 --time="$SALLOC_TIME_LIMIT" --no-shell --job-name="$RUNNER_NAME" || return 1
     JOB_ID=$(squeue --name="$RUNNER_NAME" -u "$USER" -h -o %A | head -n1)
+    [[ "$JOB_ID" =~ ^[0-9]+$ ]] || { echo "Could not resolve Slurm allocation" >&2; return 1; }
+
+    if [[ "$BENCH_SCRIPT" == *cache_sources_mtp.sh ]]; then
+        NVME_HOST_DIR=""
+        cleanup_offload_cache() {
+            local rc=$?
+            if [[ -n "$NVME_HOST_DIR" ]]; then
+                timeout --kill-after=10s 300s srun --jobid="$JOB_ID" rm -rf -- "$NVME_HOST_DIR" || rc=1
+            fi
+            scancel "$JOB_ID" || true
+            exit "$rc"
+        }
+        trap cleanup_offload_cache EXIT
+        trap 'exit 130' INT
+        trap 'exit 143' TERM
+    fi
+
+    if [[ "$KV_OFFLOADING" == *nvme* ]]; then
+        # This directory belongs only to the validated numeric allocation above.
+        export NVME_HOST_DIR="/scratch/inferencex-kv-$JOB_ID"
+        srun --jobid="$JOB_ID" mkdir -m 700 "$NVME_HOST_DIR" || return 1
+        export NVME_OFFLOAD_DIR=/kv-offload
+        CONTAINER_MOUNTS+=",$NVME_HOST_DIR:$NVME_OFFLOAD_DIR"
+    fi
 
     # Bench scripts skip `hf download` when MODEL is a local path.
     export MODEL="$MODEL_PATH"
diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index a09a238e40..dd8a4b9c72 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -180,6 +180,11 @@ rm -rf "$SRT_REPO_DIR"
 
 setup_srt_slurm "$SRT_REPO_DIR" "$FRAMEWORK" "$USES_DCGM_POWER" || exit 1
 
+if [[ "$CONFIG_FILE" == recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml ]]; then
+    # This validation needs all five DP metrics endpoints and verifies the PR overlay after setup.
+    git -C "$SRT_REPO_DIR" apply "$GITHUB_WORKSPACE/runners/srt-slurm/validation/pr56318.patch" || exit 1
+fi
+
 if [[ "$FRAMEWORK" == "dynamo-trt" && "$MODEL_PREFIX" == "dsv4" ]]; then
     SRT_SLURM_MODEL_PREFIX="deepseek-ai/DeepSeek-V4-Pro"
 fi
diff --git a/runners/srt-slurm/validation/pr56318.patch b/runners/srt-slurm/validation/pr56318.patch
new file mode 100644
index 0000000000..e8d1918038
--- /dev/null
+++ b/runners/srt-slurm/validation/pr56318.patch
@@ -0,0 +1,24 @@
+diff --git a/src/srtctl/cli/mixins/benchmark_stage.py b/src/srtctl/cli/mixins/benchmark_stage.py
+--- a/src/srtctl/cli/mixins/benchmark_stage.py
++++ b/src/srtctl/cli/mixins/benchmark_stage.py
+@@ -862,7 +862,7 @@ class BenchmarkStageMixin:
+                 env.update(self._get_aiperf_server_metrics_env())
+             elif is_custom:
+                 assert logical_endpoints is not None
+-                env.update(self._get_aiperf_server_metrics_env(logical_endpoints, logical_workers_only=True))
++                env.update(self._get_aiperf_server_metrics_env())
+         if isinstance(runner, AIPerfBenchmarkRunner) and self.config.benchmark.aiperf_package:
+             env["AIPERF_PACKAGE"] = self.config.benchmark.aiperf_package
+ 
+diff --git a/src/srtctl/cli/mixins/worker_stage.py b/src/srtctl/cli/mixins/worker_stage.py
+--- a/src/srtctl/cli/mixins/worker_stage.py
++++ b/src/srtctl/cli/mixins/worker_stage.py
+@@ -132,6 +132,8 @@ class WorkerStageMixin:
+         # Skip if dynamo.install is False (container already has dynamo installed)
+         if installs_dynamo(self.config):
+             parts.append(self.config.dynamo.get_install_commands())
++        # Fail before model load if dependency installation replaced the PR overlay.
++        parts.append("sha256sum -c /opt/pr56318-validation/validation-overlay.sha256")
+ 
+         if not parts:
+             return None
diff --git a/utils/test_agentic_offload_modes.py b/utils/test_agentic_offload_modes.py
new file mode 100644
index 0000000000..12db31ac45
--- /dev/null
+++ b/utils/test_agentic_offload_modes.py
@@ -0,0 +1,44 @@
+"""Exercise the shared offload gate with explicit recipe capabilities."""
+
+import os
+import subprocess
+from pathlib import Path
+
+import pytest
+
+LIBRARY = Path(__file__).resolve().parents[1] / "benchmarks" / "benchmark_lib.sh"
+
+
+@pytest.mark.parametrize(
+    ("mode", "capacity", "extra_modes", "expected"),
+    [
+        ("dram", "128", "", 0),
+        ("nvme", "0", "", 1),
+        ("nvme", "0", "nvme", 0),
+        ("dram+nvme", "128", "dram+nvme", 0),
+        ("dram+nvme", "0", "dram+nvme", 1),
+    ],
+)
+def test_recipe_offload_capabilities(mode, capacity, extra_modes, expected):
+    result = subprocess.run(
+        [
+            "bash",
+            "-c",
+            'source "$1"; '
+            'require_agentic_kv_offload_backend vllm-native "$2"',
+            "bash",
+            str(LIBRARY),
+            extra_modes,
+        ],
+        env={
+            **os.environ,
+            "IS_AGENTIC": "0",
+            "KV_OFFLOADING": mode,
+            "KV_OFFLOAD_BACKEND": "vllm-native",
+            "TOTAL_CPU_DRAM_GB": capacity,
+        },
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    assert result.returncode == expected, result.stderr

From 0a1bf863a33066291176269c2cf397d16548fc02 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sat, 26 Sep 2026 15:22:59 -0500
Subject: [PATCH 2/7] test: link cache-source validation changelog to PR3491

---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 04bd39c76e..a88376029a 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8973,4 +8973,4 @@
     - "Rerun vLLM PR56318 at bd57138f9b98eb64c44ea6a0f10d08c8f639820e with the four proven DeepSeek-V4-Pro hybrid cache-source points: B200 TP8 MTP3 native DRAM c8, Simple NVMe c14, and native DRAM+NVMe c14; GB300 MTP3 NIXL plus Mooncake c256."
     - "Preserve B200 GPU utilization 0.85, native DRAM utilization 0.0592, the dedicated 1 TB Simple NVMe budget, 3600-second AgentX profiles, and disabled evals. GB300 is one DEP4 prefill worker on one node plus one DEP16 decode worker across four nodes: five nodes and twenty GPUs total."
     - "Keep existing benchmark curves unchanged. Adapt the isolated old-Pro validation recipes to current InferenceX paths and current vLLM flag names, use the Python frontend, and collect all raw backend metrics with the cached-prompt source counter required."
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3491

From 1d542c78a4836cd1b62c1fd816ce05a1cfd7024f Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sat, 26 Sep 2026 15:25:15 -0500
Subject: [PATCH 3/7] test: follow relocated InferenceX test suite

---
 {utils => infx/tests}/test_agentic_offload_modes.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
 rename {utils => infx/tests}/test_agentic_offload_modes.py (94%)

diff --git a/utils/test_agentic_offload_modes.py b/infx/tests/test_agentic_offload_modes.py
similarity index 94%
rename from utils/test_agentic_offload_modes.py
rename to infx/tests/test_agentic_offload_modes.py
index 12db31ac45..63c5ac0338 100644
--- a/utils/test_agentic_offload_modes.py
+++ b/infx/tests/test_agentic_offload_modes.py
@@ -6,7 +6,7 @@
 
 import pytest
 
-LIBRARY = Path(__file__).resolve().parents[1] / "benchmarks" / "benchmark_lib.sh"
+LIBRARY = Path(__file__).resolve().parents[2] / "benchmarks" / "benchmark_lib.sh"
 
 
 @pytest.mark.parametrize(

From 56136a21fd6e4e64adee01d1e68deb498d2e7ddd Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Sat, 26 Sep 2026 16:52:30 -0400
Subject: [PATCH 4/7] refactor: migrate single-node cache-sources from bash to
 srt-slurm recipe

Replace the hand-rolled bash server script
(dsv4_fp4_b200_vllm_cache_sources_mtp.sh) with a native srt-slurm
recipe at dsv4/vllm/b200-fp4-mtp/cache-sources.yaml. Three variants
select by CONC/KV_OFFLOADING: TP8 c8 native DRAM, TP8 c14 Simple NVMe,
TP8 c14 native DRAM+NVMe. NVMe variants use host_setup to
create/teardown /scratch/inferencex-kv-$SLURM_JOB_ID and
container_mounts with {job_id} templating.

Each search-space row in nvidia-master.yaml now carries srt-recipe:,
routing to the native-single-node launch path in the b200-nscale
launcher. The bash-specific routing, /ix mount override, NVMe directory
management, and GPU_MEMORY_UTILIZATION=0.85 export are removed from
launch_b200-nscale-slurm.sh (gpu-memory-utilization is now 0.85 in the
recipe).

A new srt-slurm patch (pr56318-overlay-validation.patch) adds
pr56318-overlay-check.sh to configs/patches/, referenced by the
recipe's setup_script to verify overlay checksums before engine start.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .../dsv4_fp4_b200_vllm_cache_sources_mtp.sh   | 108 ---------------
 .../dsv4/vllm/b200-fp4-mtp/cache-sources.yaml | 127 ++++++++++++++++++
 configs/nvidia-master.yaml                    |   6 +-
 infx/tests/srt_slurm/test_srt_single_node.py  |  53 ++++++++
 runners/launch_b200-nscale-slurm.sh           |  36 +----
 .../patches/pr56318-overlay-validation.patch  |  14 ++
 6 files changed, 199 insertions(+), 145 deletions(-)
 delete mode 100644 benchmarks/single_node/agentic/dsv4_fp4_b200_vllm_cache_sources_mtp.sh
 create mode 100644 benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml
 create mode 100644 runners/srt-slurm/patches/pr56318-overlay-validation.patch

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_b200_vllm_cache_sources_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_b200_vllm_cache_sources_mtp.sh
deleted file mode 100644
index 0c6a61e9c3..0000000000
--- a/benchmarks/single_node/agentic/dsv4_fp4_b200_vllm_cache_sources_mtp.sh
+++ /dev/null
@@ -1,108 +0,0 @@
-#!/usr/bin/env bash
-set -eo pipefail
-set -x
-
-# The original DeepSeek-V4-Pro TP8/MTP3 cache-source validation recipe.
-# Preserve run 34420875287's cache pressure independently of the 0813 DSpark curve.
-source "$(dirname "$0")/../../benchmark_lib.sh"
-check_env_vars MODEL MODEL_PATH TP CONC KV_OFFLOADING KV_OFFLOAD_BACKEND
-check_env_vars TOTAL_CPU_DRAM_GB RESULT_DIR DURATION PORT GPU_MEMORY_UTILIZATION
-check_env_vars EP_SIZE DP_ATTENTION DCP_SIZE PCP_SIZE EVAL_ONLY THINKING_MODE
-
-if [[ "$TP" != 8 || "$EP_SIZE" != 1 || "$DP_ATTENTION" != false || "$DCP_SIZE" != 1 || "$PCP_SIZE" != 1 ]]; then
-    echo "Cache-source validation requires TP8 without EP, DP attention, or context parallelism" >&2
-    exit 1
-fi
-export GPU_COUNT=8
-[[ -d "$MODEL_PATH" ]] || { echo "Missing pre-staged model: $MODEL_PATH" >&2; exit 1; }
-nvidia-smi
-resolve_trace_source
-install_agentic_deps
-
-export AIPERF_SERVER_METRICS_URLS="http://localhost:$PORT/metrics"
-export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="vllm:prompt_tokens_cached_by_source"
-export VLLM_ENGINE_READY_TIMEOUT_S=3600
-export VLLM_PREFIX_CACHE_RETENTION_INTERVAL=32768
-export VLLM_USE_V2_MODEL_RUNNER=1
-export VLLM_USE_RUST_FRONTEND=0
-export VLLM_DSV4_MEGA_FP8_COMBINE=1
-export VLLM_RPC_TIMEOUT=600000
-export PYTHONHASHSEED=42
-export TORCH_CUDA_ARCH_LIST=10.0
-export PYTHONNOUSERSITE=1
-export VLLM_FLOAT32_MATMUL_PRECISION=high
-
-mkdir -p "$RESULT_DIR"
-SERVER_LOG="$RESULT_DIR/server.log"
-case "$KV_OFFLOAD_BACKEND" in
-    vllm-native)
-        require_agentic_kv_offload_backend vllm-native "dram dram+nvme"
-        SECONDARY_TIERS='[]'
-        if [[ "$KV_OFFLOADING" == dram+nvme ]]; then
-            check_env_vars NVME_OFFLOAD_DIR
-            SECONDARY_TIERS="[{\"type\":\"fs\",\"root_dir\":\"$NVME_OFFLOAD_DIR\",\"locality\":\"LOCAL\"}]"
-        fi
-        OFFLOAD_CONFIG="{\"kv_connector\":\"OffloadingConnector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"spec_name\":\"TieringOffloadingSpec\",\"cpu_bytes_to_use\":$((TOTAL_CPU_DRAM_GB * 1000000000)),\"secondary_tiers\":$SECONDARY_TIERS}}"
-        ;;
-    vllm-simple)
-        require_agentic_kv_offload_backend vllm-simple nvme
-        check_env_vars NVME_OFFLOAD_DIR
-        OFFLOAD_CONFIG="{\"kv_connector\":\"SimpleCPUOffloadConnector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"kv_offload_backend\":\"disk\",\"disk_path\":\"$NVME_OFFLOAD_DIR/cache.bin\",\"disk_capacity_bytes\":$((1000000000000 / GPU_COUNT)),\"disk_buffer_slots\":4,\"lazy_offload\":false}}"
-        ;;
-    *) echo "Unsupported validation offload backend: $KV_OFFLOAD_BACKEND" >&2; exit 1 ;;
-esac
-
-# Use the same committed golden MTP curve as the shared SRT selector.
-SPEC_CONFIG=$(python3 - <<'PY'
-import json
-import os
-from infx.srt_slurm.synthetic_acceptance import GOLDEN_DIR, golden_length
-
-spec = {"method": "mtp", "num_speculative_tokens": 3}
-if os.environ["EVAL_ONLY"] != "true":
-    spec["rejection_sample_method"] = "synthetic"
-    spec["synthetic_acceptance_length"] = golden_length(
-        "dsv4", spec, os.environ["THINKING_MODE"], GOLDEN_DIR
-    )
-print(json.dumps(spec))
-PY
-)
-MAX_NUM_SEQS=$((2 * CONC))
-CAPTURE_SIZE_LIST=()
-for ((num_seqs = 1; num_seqs <= MAX_NUM_SEQS; num_seqs++)); do
-    CAPTURE_SIZE_LIST+=("$((num_seqs * 4))")
-done
-CAPTURE_SIZE_LIST+=(100 200 300 400 500)
-CUDA_GRAPH_CAPTURE_SIZES=$(printf '%s\n' "${CAPTURE_SIZE_LIST[@]}" | sort -n -u | paste -sd, -)
-COMPILATION_CONFIG="{\"cudagraph_mode\":\"FULL_AND_PIECEWISE\",\"cudagraph_capture_sizes\":[${CUDA_GRAPH_CAPTURE_SIZES}]}"
-
-{ set +x; } 2>/dev/null
-VLLM_CMD=(
-    vllm serve "$MODEL_PATH" --served-model-name "$MODEL"
-    --host 0.0.0.0 --port "$PORT" --trust-remote-code
-    --kv-cache-dtype fp8 --block-size 256 --max-model-len 1048576
-    --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION"
-    --numa-bind --enable-cumem-allocator --no-enable-flashinfer-autotune
-    --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4
-    --enable-auto-tool-choice --reasoning-parser deepseek_v4
-    --attention-config '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"indexer_kv_dtype":"mxfp4"}'
-    --speculative-config "$SPEC_CONFIG" --no-disable-hybrid-kv-cache-manager
-    --disable-uvicorn-access-log --compilation-config "$COMPILATION_CONFIG"
-    --max-num-seqs "$MAX_NUM_SEQS" --tensor-parallel-size "$TP" --data-parallel-size 1
-    --kv-transfer-config "$OFFLOAD_CONFIG"
-)
-printf '%q ' "${VLLM_CMD[@]}" | tee "$RESULT_DIR/vllm_command.txt"
-printf '\n' | tee -a "$RESULT_DIR/vllm_command.txt"
-sha256sum -c /opt/pr56318-validation/validation-overlay.sha256 | tee "$RESULT_DIR/overlay-verification.log"
-"${VLLM_CMD[@]}" > "$SERVER_LOG" 2>&1 &
-SERVER_PID=$!
-wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
-
-if [[ "$EVAL_ONLY" == true ]]; then
-    run_eval --port "$PORT"
-else
-    build_replay_cmd "$RESULT_DIR"
-    curl --fail --silent --show-error "$AIPERF_SERVER_METRICS_URLS" > "$RESULT_DIR/metrics-before.prom"
-    run_agentic_replay_and_write_outputs "$RESULT_DIR"
-    curl --fail --silent --show-error "$AIPERF_SERVER_METRICS_URLS" > "$RESULT_DIR/metrics-after.prom"
-fi
diff --git a/benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml b/benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml
new file mode 100644
index 0000000000..9a0af1a64d
--- /dev/null
+++ b/benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml
@@ -0,0 +1,127 @@
+# PR56318 cache-source validation: DeepSeek-V4-Pro TP8/MTP3 on B200 with
+# three KV offload modes (native DRAM, Simple NVMe, native DRAM+NVMe).
+# One variant per search-space row.
+base:
+  schema: 2
+  name: dsv4-fp4-b200-vllm-cache-sources
+  model:
+    path: hf:deepseek-ai/DeepSeek-V4-Pro
+    container: cquil11/vllm-cache-sources@sha256:6f9e476f6a1e1c25ac9bc1a863329538db465812fb4ac18c6553e988eae09a56
+    precision: fp4
+  resources:
+    gpu_type: b200
+    gpus_per_node: 8
+  frontend:
+    type: vllm
+    enable_multiple_frontends: false
+  observability:
+    enabled: false
+    tachometer:
+      enabled: false
+  engine:
+    type: vllm
+    connector: null
+  health_check:
+    interval_seconds: 10
+    max_attempts: 720
+  setup_script: pr56318-overlay-check.sh
+  roles:
+    agg:
+      nodes: 1
+      workers: 1
+      args:
+        served-model-name: deepseek-ai/DeepSeek-V4-Pro
+        trust-remote-code: true
+        kv-cache-dtype: fp8
+        block-size: 256
+        max-model-len: 1048576
+        gpu-memory-utilization: 0.85
+        numa-bind: true
+        enable-cumem-allocator: true
+        no-enable-flashinfer-autotune: true
+        tokenizer-mode: deepseek_v4
+        tool-call-parser: deepseek_v4
+        enable-auto-tool-choice: true
+        reasoning-parser: deepseek_v4
+        attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"indexer_kv_dtype":"mxfp4"}'
+        speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
+        no-disable-hybrid-kv-cache-manager: true
+        disable-uvicorn-access-log: true
+        tensor-parallel-size: 8
+        data-parallel-size: 1
+      env:
+        VLLM_ENGINE_READY_TIMEOUT_S: '3600'
+        VLLM_PREFIX_CACHE_RETENTION_INTERVAL: '32768'
+        VLLM_USE_V2_MODEL_RUNNER: '1'
+        VLLM_USE_RUST_FRONTEND: '0'
+        VLLM_DSV4_MEGA_FP8_COMBINE: '1'
+        VLLM_RPC_TIMEOUT: '600000'
+        PYTHONHASHSEED: '42'
+        TORCH_CUDA_ARCH_LIST: '10.0'
+        PYTHONNOUSERSITE: '1'
+        VLLM_FLOAT32_MATMUL_PRECISION: high
+  benchmark:
+    type: custom
+    command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+    env:
+      MODEL: deepseek-ai/DeepSeek-V4-Pro
+      AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'vllm:prompt_tokens_cached_by_source'
+
+# ---------- Variant 1: TP8 c8 native DRAM ----------
+# max-num-seqs = 2 * CONC = 16; capture sizes = {4*n for n=1..16} ∪ {100..500}
+override_tp8_c8_dram_native:
+  roles:
+    agg:
+      gpus: 8
+      args:
+        max-num-seqs: 16
+        compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,100,200,300,400,500]}'
+        kv-transfer-config: '{"kv_connector":"OffloadingConnector","kv_role":"kv_both","kv_connector_extra_config":{"spec_name":"TieringOffloadingSpec","cpu_bytes_to_use":128000000000,"secondary_tiers":[]}}'
+  benchmark:
+    env:
+      CONC: '8'
+      KV_OFFLOADING: dram
+      TOTAL_CPU_DRAM_GB: '128'
+
+# ---------- Variant 2: TP8 c14 Simple NVMe ----------
+# max-num-seqs = 2 * CONC = 28; disk_capacity_bytes = 1e12 / 8
+override_tp8_c14_nvme_simple:
+  container_mounts:
+    "/scratch/inferencex-kv-{job_id}": "/kv-offload"
+  host_setup:
+    commands: ["mkdir -m 700 /scratch/inferencex-kv-$SLURM_JOB_ID"]
+    teardown: ["rm -rf /scratch/inferencex-kv-$SLURM_JOB_ID"]
+    nodes: workers
+  roles:
+    agg:
+      gpus: 8
+      args:
+        max-num-seqs: 28
+        compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96,100,104,108,112,200,300,400,500]}'
+        kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"kv_offload_backend":"disk","disk_path":"/kv-offload/cache.bin","disk_capacity_bytes":125000000000,"disk_buffer_slots":4,"lazy_offload":false}}'
+  benchmark:
+    env:
+      CONC: '14'
+      KV_OFFLOADING: nvme
+
+# ---------- Variant 3: TP8 c14 native DRAM + NVMe ----------
+# max-num-seqs = 2 * CONC = 28; cpu_bytes = 128e9, fs tier at /kv-offload
+override_tp8_c14_dramnvme_native:
+  container_mounts:
+    "/scratch/inferencex-kv-{job_id}": "/kv-offload"
+  host_setup:
+    commands: ["mkdir -m 700 /scratch/inferencex-kv-$SLURM_JOB_ID"]
+    teardown: ["rm -rf /scratch/inferencex-kv-$SLURM_JOB_ID"]
+    nodes: workers
+  roles:
+    agg:
+      gpus: 8
+      args:
+        max-num-seqs: 28
+        compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96,100,104,108,112,200,300,400,500]}'
+        kv-transfer-config: '{"kv_connector":"OffloadingConnector","kv_role":"kv_both","kv_connector_extra_config":{"spec_name":"TieringOffloadingSpec","cpu_bytes_to_use":128000000000,"secondary_tiers":[{"type":"fs","root_dir":"/kv-offload","locality":"LOCAL"}]}}'
+  benchmark:
+    env:
+      CONC: '14'
+      KV_OFFLOADING: dram+nvme
+      TOTAL_CPU_DRAM_GB: '128'
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 30d7264832..30a355c896 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8398,12 +8398,12 @@ dsv4-fp4-b200-vllm-agentic-cache-sources-mtp:
     agentic-coding:
     - dram-utilization: 0.0592
       search-space:
-      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: vllm-native }, spec-decoding: mtp, conc-list: [8] }
+      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: vllm-native }, spec-decoding: mtp, conc-list: [8], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml }
     - search-space:
-      - { tp: 8, kv-offloading: nvme, kv-offload-backend: { name: vllm-simple }, spec-decoding: mtp, conc-list: [14] }
+      - { tp: 8, kv-offloading: nvme, kv-offload-backend: { name: vllm-simple }, spec-decoding: mtp, conc-list: [14], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml }
     - dram-utilization: 0.0592
       search-space:
-      - { tp: 8, kv-offloading: [dram, nvme], kv-offload-backend: { name: vllm-native }, spec-decoding: mtp, conc-list: [14] }
+      - { tp: 8, kv-offloading: [dram, nvme], kv-offload-backend: { name: vllm-native }, spec-decoding: mtp, conc-list: [14], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml }
 
 dsv4-fp4-gb300-dynamo-vllm-agentic-cache-sources-mtp-disagg:
   image: cquil11/vllm-cache-sources@sha256:6ce87fcb9ef8b929f7a8270bc6fbba3c219c3440cd77d3bf1ae74db5a1cb5715
diff --git a/infx/tests/srt_slurm/test_srt_single_node.py b/infx/tests/srt_slurm/test_srt_single_node.py
index d0b160dd5f..2f77ec2c9d 100644
--- a/infx/tests/srt_slurm/test_srt_single_node.py
+++ b/infx/tests/srt_slurm/test_srt_single_node.py
@@ -453,3 +453,56 @@ def test_b300_keeps_agentic_and_explicit_collector_dispatch(tmp_path, collector)
     assert calls[-1][-2:] == ["bash", expected]
     assert "--jobid=42" in calls[-1]
     assert (tmp_path / "cancelled").read_text() == "42\n"
+
+
+# ---------------------------------------------------------------------------
+# Cache-sources recipe: each variant renders the expected kv-transfer-config
+# and capture sizes.
+# ---------------------------------------------------------------------------
+
+CACHE_SOURCES_RECIPE = str(ROOT / "benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml")
+
+CACHE_SOURCES_BASE_ENV = {
+    "MODEL": "deepseek-ai/DeepSeek-V4-Pro",
+    "IMAGE": "cquil11/vllm-cache-sources@sha256:6f9e476f6a1e1c25ac9bc1a863329538db465812fb4ac18c6553e988eae09a56",
+    "PRECISION": "fp4", "FRAMEWORK": "vllm", "TP": "8", "EP_SIZE": "1",
+    "DP_ATTENTION": "false", "PP_SIZE": "1", "DCP_SIZE": "1", "PCP_SIZE": "1",
+    "GPU_COUNT": "8", "IS_AGENTIC": "1", "SPEC_DECODING": "mtp",
+    "THINKING_MODE": "off", "EVAL_ONLY": "false", "RUN_EVAL": "false",
+    "MODEL_PREFIX": "dsv4", "DURATION": "3600", "RESULT_FILENAME": "test",
+    "GPU_MONITOR_INTERVAL": "5",
+}
+
+
+@pytest.mark.parametrize(
+    ("conc", "kv_offloading", "total_dram", "variant_suffix", "connector", "kv_path"),
+    [
+        ("8", "dram", "128", "c8_dram_native", "OffloadingConnector", None),
+        ("14", "nvme", "0", "c14_nvme_simple", "SimpleCPUOffloadConnector", "/kv-offload/cache.bin"),
+        ("14", "dram+nvme", "128", "c14_dramnvme_native", "OffloadingConnector", "/kv-offload"),
+    ],
+)
+def test_cache_sources_variant_renders_expected_offload(
+    conc, kv_offloading, total_dram, variant_suffix, connector, kv_path,
+):
+    env = {**CACHE_SOURCES_BASE_ENV, "CONC": conc, "KV_OFFLOADING": kv_offloading,
+           "TOTAL_CPU_DRAM_GB": total_dram}
+    config, recipe = select_recipe(CACHE_SOURCES_RECIPE, env)
+    assert config.endswith(f"override_tp8_{variant_suffix}")
+
+    transfer = json.loads(recipe["roles"]["agg"]["args"]["kv-transfer-config"])
+    assert transfer["kv_connector"] == connector
+
+    max_num_seqs = recipe["roles"]["agg"]["args"]["max-num-seqs"]
+    assert max_num_seqs == 2 * int(conc)
+
+    compilation = json.loads(recipe["roles"]["agg"]["args"]["compilation-config"])
+    expected_captures = sorted(set([4 * n for n in range(1, max_num_seqs + 1)] + [100, 200, 300, 400, 500]))
+    assert compilation["cudagraph_capture_sizes"] == expected_captures
+
+    if kv_path and connector == "SimpleCPUOffloadConnector":
+        assert transfer["kv_connector_extra_config"]["disk_path"] == kv_path
+    elif kv_path and connector == "OffloadingConnector":
+        tiers = transfer["kv_connector_extra_config"]["secondary_tiers"]
+        assert len(tiers) == 1
+        assert tiers[0]["root_dir"] == kv_path
diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
index 291538f32b..1c6c7afc04 100755
--- a/runners/launch_b200-nscale-slurm.sh
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -786,10 +786,6 @@ run_agentic() {
     if [[ ! -f "$BENCH_SCRIPT" ]]; then
         BENCH_SCRIPT="${BENCH_BASE}${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh"
     fi
-    if [[ "$MODEL" == "deepseek-ai/DeepSeek-V4-Pro" && "$FRAMEWORK" == vllm && "$SPEC_DECODING" == mtp && "$IMAGE" == cquil11/vllm-cache-sources@sha256:6f9e476f6a1e1c25ac9bc1a863329538db465812fb4ac18c6553e988eae09a56 ]]; then
-        BENCH_SCRIPT="benchmarks/single_node/agentic/dsv4_fp4_b200_vllm_cache_sources_mtp.sh"
-        export GPU_MEMORY_UTILIZATION=0.85
-    fi
     LOCK_FILE="${SQUASH_FILE}.lock"
 
     # TODO(Cam): lmsysorg/sglang:deepseek-v4-blackwell installs sglang editable at
@@ -797,7 +793,7 @@ run_agentic() {
     # the default $GITHUB_WORKSPACE:/workspace/ bind-mount masks the install and
     # breaks `import sglang`. Mount this one image at /ix instead; drop the
     # conditional once the image stops installing editable under /workspace.
-    if [[ "$IMAGE" == *deepseek-v4-blackwell* || "$BENCH_SCRIPT" == *cache_sources_mtp.sh ]]; then
+    if [[ "$IMAGE" == *deepseek-v4-blackwell* ]]; then
         CONTAINER_MOUNT_DIR=/ix
     else
         CONTAINER_MOUNT_DIR=/workspace
@@ -814,37 +810,9 @@ run_agentic() {
     else
         CONTAINER_MOUNTS="$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR,$MODEL_PATH:$MODEL_PATH,$AIPERF_MMAP_CACHE_HOST_PATH:/aiperf_mmap_cache"
     fi
-    if [[ "$BENCH_SCRIPT" == *cache_sources_mtp.sh ]]; then
-        export INFMAX_CONTAINER_WORKSPACE="$CONTAINER_MOUNT_DIR"
-        export RESULT_DIR="$CONTAINER_MOUNT_DIR/results"
-    fi
 
-    salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --gres=gpu:$GPU_COUNT --exclusive --mem=0 --time="$SALLOC_TIME_LIMIT" --no-shell --job-name="$RUNNER_NAME" || return 1
+    salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --gres=gpu:$GPU_COUNT --exclusive --mem=0 --time="$SALLOC_TIME_LIMIT" --no-shell --job-name="$RUNNER_NAME"
     JOB_ID=$(squeue --name="$RUNNER_NAME" -u "$USER" -h -o %A | head -n1)
-    [[ "$JOB_ID" =~ ^[0-9]+$ ]] || { echo "Could not resolve Slurm allocation" >&2; return 1; }
-
-    if [[ "$BENCH_SCRIPT" == *cache_sources_mtp.sh ]]; then
-        NVME_HOST_DIR=""
-        cleanup_offload_cache() {
-            local rc=$?
-            if [[ -n "$NVME_HOST_DIR" ]]; then
-                timeout --kill-after=10s 300s srun --jobid="$JOB_ID" rm -rf -- "$NVME_HOST_DIR" || rc=1
-            fi
-            scancel "$JOB_ID" || true
-            exit "$rc"
-        }
-        trap cleanup_offload_cache EXIT
-        trap 'exit 130' INT
-        trap 'exit 143' TERM
-    fi
-
-    if [[ "$KV_OFFLOADING" == *nvme* ]]; then
-        # This directory belongs only to the validated numeric allocation above.
-        export NVME_HOST_DIR="/scratch/inferencex-kv-$JOB_ID"
-        srun --jobid="$JOB_ID" mkdir -m 700 "$NVME_HOST_DIR" || return 1
-        export NVME_OFFLOAD_DIR=/kv-offload
-        CONTAINER_MOUNTS+=",$NVME_HOST_DIR:$NVME_OFFLOAD_DIR"
-    fi
 
     # Bench scripts skip `hf download` when MODEL is a local path.
     export MODEL="$MODEL_PATH"
diff --git a/runners/srt-slurm/patches/pr56318-overlay-validation.patch b/runners/srt-slurm/patches/pr56318-overlay-validation.patch
new file mode 100644
index 0000000000..6cfa15c934
--- /dev/null
+++ b/runners/srt-slurm/patches/pr56318-overlay-validation.patch
@@ -0,0 +1,14 @@
+diff --git a/configs/patches/pr56318-overlay-check.sh b/configs/patches/pr56318-overlay-check.sh
+new file mode 100644
+index 000000000..000000001
+--- /dev/null
++++ b/configs/patches/pr56318-overlay-check.sh
+@@ -0,0 +1,8 @@
++#!/bin/bash
++# Verify PR56318 validation overlay checksums before engine start.
++# A no-op when run on images that do not carry the overlay.
++set -euo pipefail
++OVERLAY=/opt/pr56318-validation/validation-overlay.sha256
++if [ -f "$OVERLAY" ]; then
++    sha256sum -c "$OVERLAY"
++fi

From ef523df3b3ccb2f51d5ca385804857156fbbc1c3 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 27 Sep 2026 01:31:19 -0500
Subject: [PATCH 5/7] fix: use current AgentX entrypoint in cache-source
 validation

---
 .../gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml     | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml
index 853ca3f404..baf5f7e97e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml
@@ -214,7 +214,7 @@ srun_options:
 
 benchmark:
   type: custom
-  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  command: bash /infmax-workspace/benchmarks/srt_agentic.sh
   env:
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:prompt_tokens_cached_by_source"
     INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"

From 968e80fe593a91e71aa014acbc9b05e78e8087dc Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 1 Oct 2026 12:30:05 -0500
Subject: [PATCH 6/7] test: rerun cache-source sweeps with bounded-range sweep
 accounting

---
 .../agentx/cache-sources-dep4-dep16-c256-mtp.yaml     |  4 ++--
 .../dsv4/vllm/b200-fp4-mtp/cache-sources.yaml         |  2 +-
 inferencex-e2e/configs/nvidia-master.yaml             |  4 ++--
 inferencex-e2e/perf-changelog.yaml                    | 11 +++++++++++
 4 files changed, 16 insertions(+), 5 deletions(-)

diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml
index baf5f7e97e..41aa39a9a9 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml
@@ -9,14 +9,14 @@ name: "dsv4-vllm-cache-sources-gb300-dep4-dep16-c256"
 
 model:
   path: "deepseek-v4-pro"
-  container: "cquil11/vllm-cache-sources@sha256:6ce87fcb9ef8b929f7a8270bc6fbba3c219c3440cd77d3bf1ae74db5a1cb5715"
+  container: "cquil11/vllm-cache-sources@sha256:188445e9835a0ed9cd022f0e70c61dd3fd9d24792008841d4747da1797e7fe4b"
   precision: "fp4"
 
 identity:
   model:
     repo: "deepseek-ai/DeepSeek-V4-Pro"
   container:
-    image: "cquil11/vllm-cache-sources@sha256:6ce87fcb9ef8b929f7a8270bc6fbba3c219c3440cd77d3bf1ae74db5a1cb5715"
+    image: "cquil11/vllm-cache-sources@sha256:188445e9835a0ed9cd022f0e70c61dd3fd9d24792008841d4747da1797e7fe4b"
   frameworks:
     dynamo: "1.4.0"
 
diff --git a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml
index 9a0af1a64d..701a48e8d2 100644
--- a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml
+++ b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml
@@ -6,7 +6,7 @@ base:
   name: dsv4-fp4-b200-vllm-cache-sources
   model:
     path: hf:deepseek-ai/DeepSeek-V4-Pro
-    container: cquil11/vllm-cache-sources@sha256:6f9e476f6a1e1c25ac9bc1a863329538db465812fb4ac18c6553e988eae09a56
+    container: cquil11/vllm-cache-sources@sha256:873b8bcb4cc734cfc3e3b7f9d2c1495be329439a944f76a7b356448be6a61641
     precision: fp4
   resources:
     gpu_type: b200
diff --git a/inferencex-e2e/configs/nvidia-master.yaml b/inferencex-e2e/configs/nvidia-master.yaml
index 7bb01fedc9..5b8d2710e3 100644
--- a/inferencex-e2e/configs/nvidia-master.yaml
+++ b/inferencex-e2e/configs/nvidia-master.yaml
@@ -8385,7 +8385,7 @@ dsv41flash-fp4-b200-sglang-agentic-dspark:
 
 # PR56318: rerun the four successful hybrid cache-source validation points.
 dsv4-fp4-b200-vllm-agentic-cache-sources-mtp:
-  image: cquil11/vllm-cache-sources@sha256:6f9e476f6a1e1c25ac9bc1a863329538db465812fb4ac18c6553e988eae09a56
+  image: cquil11/vllm-cache-sources@sha256:873b8bcb4cc734cfc3e3b7f9d2c1495be329439a944f76a7b356448be6a61641
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:b200-nscale
@@ -8404,7 +8404,7 @@ dsv4-fp4-b200-vllm-agentic-cache-sources-mtp:
       - { tp: 8, kv-offloading: [dram, nvme], kv-offload-backend: { name: vllm-native }, spec-decoding: mtp, conc-list: [14], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/b200-fp4-mtp/cache-sources.yaml }
 
 dsv4-fp4-gb300-dynamo-vllm-agentic-cache-sources-mtp-disagg:
-  image: cquil11/vllm-cache-sources@sha256:6ce87fcb9ef8b929f7a8270bc6fbba3c219c3440cd77d3bf1ae74db5a1cb5715
+  image: cquil11/vllm-cache-sources@sha256:188445e9835a0ed9cd022f0e70c61dd3fd9d24792008841d4747da1797e7fe4b
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:gb300-nv
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index 0db4cf1f9a..bd6d05d255 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9192,3 +9192,14 @@
     - "Preserve B200 GPU utilization 0.85, native DRAM utilization 0.0592, the dedicated 1 TB Simple NVMe budget, 3600-second AgentX profiles, and disabled evals. GB300 is one DEP4 prefill worker on one node plus one DEP16 decode worker across four nodes: five nodes and twenty GPUs total."
     - "Keep existing benchmark curves unchanged. Adapt the isolated old-Pro validation recipes to current InferenceX paths and current vLLM flag names, use the Python frontend, and collect all raw backend metrics with the cached-prompt source counter required."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3491
+
+- config-keys:
+    - dsv4-fp4-b200-vllm-agentic-cache-sources-mtp
+    - dsv4-fp4-gb300-dynamo-vllm-agentic-cache-sources-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  no-evals: true
+  description:
+    - "Rerun the four DeepSeek-V4-Pro cache-source points with vLLM PR56318 commit 6ce00ee2ebcf29076d671786c36bf8f8ca71bd0b: B200 native DRAM c8, Simple NVMe c14, native DRAM+NVMe c14, and GB300 NIXL plus Mooncake Store c256."
+    - "Keep the prior runtimes and serving settings, 3600-second AIPerf AgentX profiles, Python frontend, and Prometheus source collection. Update the attribution overlay for bounded attention ranges and sweep-line accounting."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3491

From 0efad80962a04adbefe8e2e32d372b5520eeb22d Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 1 Oct 2026 12:52:50 -0500
Subject: [PATCH 7/7] fix: preserve B200 DeepSeek Pro validation checkpoint

---
 .../infx/launch/drivers/srt/checkout.py       |  4 +-
 .../infx/launch/drivers/srt/models.py         | 10 +++++
 .../infx/tests/launch/test_srt_policy.py      | 38 +++++++++++++++++++
 inferencex-e2e/perf-changelog.yaml            | 10 +++++
 4 files changed, 61 insertions(+), 1 deletion(-)

diff --git a/inferencex-e2e/infx/launch/drivers/srt/checkout.py b/inferencex-e2e/infx/launch/drivers/srt/checkout.py
index d7c5e12a48..42c030fb39 100644
--- a/inferencex-e2e/infx/launch/drivers/srt/checkout.py
+++ b/inferencex-e2e/infx/launch/drivers/srt/checkout.py
@@ -87,7 +87,9 @@ def prepare_checkout(run: SrtRun, destination: Path, *, power: bool) -> Checkout
     if run.request.config_file == (
         "recipes/dsv4/vllm/gb300-fp4/agentx/cache-sources-dep4-dep16-c256-mtp.yaml"
     ):
-        _git("-C", destination, "apply", run.workspace / "runners/srt-slurm/validation/pr56318.patch")
+        _git(
+            "-C", destination, "apply", run.workspace / "runners/srt-slurm/validation/pr56318.patch"
+        )
     head = _git("-C", destination, "rev-parse", "HEAD", capture=True)
     if head != commit:
         raise LaunchError(f"srt-slurm checkout is at {head}, expected {commit}")
diff --git a/inferencex-e2e/infx/launch/drivers/srt/models.py b/inferencex-e2e/infx/launch/drivers/srt/models.py
index bb7f701534..63ff070d23 100644
--- a/inferencex-e2e/infx/launch/drivers/srt/models.py
+++ b/inferencex-e2e/infx/launch/drivers/srt/models.py
@@ -39,6 +39,16 @@ class Override:
 
 OVERRIDES: dict[str, tuple[Override, ...]] = {
     "b200-nscale": (
+        Override(
+            Match(
+                any_of("dsv4"),
+                any_of("fp4"),
+                any_of("vllm"),
+                multinode=False,
+                model_glob="deepseek-ai/DeepSeek-V4-Pro",
+            ),
+            entry="DeepSeek-V4-Pro-NVFP4",
+        ),
         Override(
             Match(any_of("glm5.1"), frameworks=any_of("tilert")),
             entry="GLM-5.1-FP8@shared",
diff --git a/inferencex-e2e/infx/tests/launch/test_srt_policy.py b/inferencex-e2e/infx/tests/launch/test_srt_policy.py
index cbaf6d97da..4893549428 100644
--- a/inferencex-e2e/infx/tests/launch/test_srt_policy.py
+++ b/inferencex-e2e/infx/tests/launch/test_srt_policy.py
@@ -106,6 +106,44 @@ def test_a_checkpoint_that_must_be_readable_fails_before_submission(tmp_path, mo
     assert host_path(c, checkpoint(c, request(MODEL="org/S"))) == tmp_path / "shared/s"
 
 
+@pytest.mark.parametrize(
+    ("model", "framework", "multinode", "directory"),
+    [
+        ("DeepSeek-V4-Pro", "vllm", "false", "converted"),
+        ("DeepSeek-V4-Pro-0813", "vllm", "false", "august"),
+        ("DeepSeek-V4-Pro", "vllm", "true", "original"),
+        ("DeepSeek-V4-Pro", "sglang", "false", "original"),
+    ],
+)
+def test_b200_original_pro_single_node_vllm_keeps_converted_checkpoint(
+    tmp_path,
+    model,
+    framework,
+    multinode,
+    directory,
+):
+    c = cluster(tmp_path)
+    c.bind_id("b200-nscale")
+    entry = c.models.entries["M"]
+    entries = {
+        name: entry.model_copy(update={"dir": path})
+        for name, path in {
+            "DeepSeek-V4-Pro": "original",
+            "DeepSeek-V4-Pro-NVFP4": "converted",
+            "DeepSeek-V4-Pro-0813": "august",
+        }.items()
+    }
+    c = c.model_copy(update={"models": c.models.model_copy(update={"entries": entries})})
+    point = request(
+        MODEL=f"deepseek-ai/{model}",
+        MODEL_PREFIX="dsv4",
+        PRECISION="fp4",
+        FRAMEWORK=framework,
+        IS_MULTINODE=multinode,
+    )
+    assert single_node_model_path(c, point) == str(tmp_path / "shared" / directory)
+
+
 def test_a_points_own_model_path_is_what_its_job_serves(tmp_path, monkeypatch):
     c = cluster(tmp_path)
     host = request(MODEL="org/M", MODEL_PATH="/host/m")
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index bd6d05d255..31e35e46e7 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9203,3 +9203,13 @@
     - "Rerun the four DeepSeek-V4-Pro cache-source points with vLLM PR56318 commit 6ce00ee2ebcf29076d671786c36bf8f8ca71bd0b: B200 native DRAM c8, Simple NVMe c14, native DRAM+NVMe c14, and GB300 NIXL plus Mooncake Store c256."
     - "Keep the prior runtimes and serving settings, 3600-second AIPerf AgentX profiles, Python frontend, and Prometheus source collection. Update the attribution overlay for bounded attention ranges and sweep-line accounting."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3491
+
+- config-keys:
+    - dsv4-fp4-b200-vllm-agentic-cache-sources-mtp
+    - dsv4-fp4-gb300-dynamo-vllm-agentic-cache-sources-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  no-evals: true
+  description:
+    - "Restore the original B200 single-node vLLM DeepSeek-V4-Pro-NVFP4 checkpoint mapping after the launcher migration. Repeat the four cache-source validation points with unchanged images and serving settings."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3491